Your first recall
Once your client is connected and admitted to a worktree, the core loop is two tool calls: capture a belief, then retrieve it.
1. Capture — zephr_remember
zephr_remember captures an unreviewed belief with provenance, or uses the local review
capability to confirm or reject its current open version. An agent records what it has learned
with the source and anchor it came from.
2. Retrieve — zephr_recall
zephr_recall retrieves eligible belief versions via hybrid lexical + semantic search
(SQLite FTS5 for lexical, sqlite-vec for semantic, fused by reciprocal rank). It returns
ranked items with a final score and contradiction belief IDs.
3. Explain — zephr_why
zephr_why explains a belief version: its content, kind, review state, validity, version
count, worktree-scoped relationships, and provenance when recorded.
What you should observe
- Recalled beliefs carry a provenance chain, not bare text.
- Recall surfaces only eligible states — by default confirmed and valid beliefs — and abstains rather than guessing when evidence is missing, stale, or contradictory.
- The Trust Firewall (
zephr_rules) can only narrow retrieval, never broaden it.
Smoke test without persistence
If you omit ZEPHR_DB_PATH, the server falls back to an in-memory store — fine for a smoke
test, but everything is discarded when the process exits. Use a real file path for any work
you want to keep.
Next
Read Beliefs to understand review states and version chains, or the MCP tools reference for the quick-start subset and generated inventory.